Le Chonk

In the episode's first major segment, the host discusses the launch of Reflection's new AI model, Beam, which is designed to be a competitive open-weight alternative in the AI space. Reflection, a New York City-based company backed by Nvidia, announced Beam as an open model that rivals GLM 5.2 in reasoning capabilities while requiring significantly less compute—three to four times less—and approaching Qwen 3.8-Max on agentic tasks. The model is aimed at business and developer audiences for coding and AI agent testing, and it is built to perform efficiently at a cost advantage. Reflection plans to release detailed technical data later this month, highlighting that Beam contains about 500 billion parameters, roughly half the size of some other models, but with performance benefits in reasoning and coding. The company's approach involved training Beam on a massive dataset of 23.8 trillion tokens, using thousands of GPUs, and optimizing its architecture through several stages, including a reinforcement learning phase.

The episode then shifts to an analysis segment on AI subscription pricing, discussing how consumers and small businesses are paying for AI services. SemiAnalysis reports that Claude's plans from Anthropic offer roughly five times more API-value per dollar compared to OpenAI's GPT plans—specifically GPT-6.1. The host explains the mechanics: subscription plans are based on credits, which vary in cost depending on the workload and model used, making direct dollar comparisons complex. It is noted that Anthropic's plans are more generous in limits than OpenAI's, which has recently reduced the API-equivalent value for certain tiers. The episode emphasizes that Opus 5.5 from Claude offers significantly better value than GPT-6.1 from OpenAI, even when considering differences like maximum session length. This disparity is expected to grow as models become cheaper and more advanced, with OpenAI possibly adjusting its approach to limits.

Next, the conversation turns to corporate AI use, revealing that Meta and Microsoft are scaling back their internal use of Anthropic's Claude models. Meta's internal usage of Claude-based tools, including 'cloud code,' has dropped from around 60,000 users earlier this year to approximately 30,000, partly due to layoffs in Meta and partly due to Meta developing its own AI models like 'Mews Spark.' Microsoft has similarly reduced its internal use, with projections indicating it would spend over a billion dollars on Anthropic's AI technology earlier in the year, but this has now been scaled back by over a third. Microsoft is shifting toward its internally developed tools, reducing reliance on external models. These cuts are part of broader cost-saving efforts amidst rising AI costs and the strategic move by Big Tech companies to develop their own AI infrastructure.

The host discusses the release of Mistral's latest model, called 'Le Chonk,' a 1-trillion-parameter open-weight model aimed at general-agentic tasks, trained on 4,000 Nvidia Grace Blackwell GPUs. Mistral positions Le Chonk as a competitor to both closed and open models, claiming it offers performance comparable to the best closed models while being more cost-effective to train. The model is designed with cybersecurity in mind, emphasizing the benefits of open weights for building and deploying defense mechanisms without the risk of sudden deprecation. Beyond cybersecurity, Le Chonk excels in coding, domain-specific work, and multi-modality. Mistral also highlighted their own infrastructure and expertise in training, noting they will release the model's weights on October 27th, making it accessible for enterprise deployment either on-premise or via their API.

The final segment covers the viral phenomenon of AI-generated cartoons mimicking New Yorker cartoonists. ChatGPT was used to generate cartoons in the style of famous contributors, and these images were then signed with real cartoonists' pen names, like Brendan Lopper and Emily Flake, without their approval. Lopper expressed feeling that this was a violation of personal rights, as his signature was faked, and overall it raised concerns about AI training data and attribution. The cartoons with signatures, some of whom are legendary cartoonists, circulated widely on social media, causing a mix of amusement and discomfort. The episode illustrates both the creative potential and ethical challenges of AI in generating personalized art and signatures.